EnvFactory: Scaling Tool-Use Agents via Executable Environments Synthesis and Robust RL

EnvFactory is a fully automated framework that discovers and verifies executable tool-use environments from real sources, then generates natural multi-turn trajectories for agentic RL.

Across 85 environments in 7 domains, it produced 2,575 training trajectories and improved Qwen3 models on several tool-use and conversational benchmarks.

The method delivers strong results despite using fewer environments than prior approaches, including gains on BFCLv3, MCP-Atlas, τ^2-Bench, and VitaBench.

Overall, EnvFactory cuts the manual effort and cost of training tool-using LLM agents while improving performance across domains.